Heterogeneous service node quality perception and reward distribution method for alliance computing power network

Abstract

The application provides a heterogeneous service node quality perception and reward distribution method for a consortium computing power network, comprising: receiving or responding to a computing task request, selecting a candidate computing node; obtaining the service quality and response time delay of each candidate computing node, and constructing a node evaluation index; using an improved confidence upper bound algorithm to evaluate the quality of the candidate computing node, selecting a preset number of candidate computing nodes as the target of maximizing the quality of the computing node; identifying the association relationship of the candidate computing node to divide multiple alliances; controlling the candidate computing nodes to cooperatively execute a distributed computing task, and using a multi-agent reinforcement learning-based model to determine the computing contribution of each candidate computing node in each time slot; and calculating a resource incentive value through a two-level distribution mechanism to update the scheduling priority of each candidate computing node. The application can improve the throughput and stability of a heterogeneous distributed computing system.

Description

Technical Field

[0001] This invention belongs to the field of distributed computing technology, specifically relating to a method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing power network. Background Technology

[0002] The rapid development of large language models and deep learning applications in artificial intelligence has created unprecedented demands for computing resources. This demand has driven the emergence of distributed computing platforms, which provide services by aggregating heterogeneous computing resources from different providers. The shift from centralized to decentralized infrastructure has improved resource utilization, but it has also brought challenges arising from node heterogeneity.

[0003] Existing technologies for solving the node selection problem mostly assume that node performance is static or known, or they only use simple multi-armed slot machine algorithms to optimize a single metric (such as focusing only on computing power or only on network performance). These existing technologies ignore the fact that the heterogeneity of participating nodes can cause significant differences in two independent dimensions: hardware service quality and network response latency. They cannot accurately characterize the node evaluation metrics that affect the overall system performance. Furthermore, when determining the computing contribution of nodes and allocating resources, existing technologies often assume that computing nodes interact completely independently. However, in real network environments, nodes usually self-organize into alliance structures based on geographical regions or trust protocols. Existing simple average allocation schemes ignore this hierarchical physical alliance structure and the hardware resource quota constraints faced by each underlying node. As a result, nodes with high-performance hardware often do not receive matching resource incentives and scheduling priority biases, thereby reducing their willingness to collaboratively provide computing power. Ultimately, this leads to a decrease in the overall computing efficiency of the distributed cluster and system instability. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network, so as to solve the problems of suboptimal node selection, unreasonable resource scheduling and configuration, and low overall computing throughput efficiency of the cluster when facing heterogeneous computing nodes in existing distributed computing systems. This is due to the neglect of the combined effects of hardware service quality and network latency, as well as the lack of consideration for the underlying physical consortium structure and hardware resource quota constraints of the nodes.

[0005] This invention provides a method for quality awareness and reward allocation of heterogeneous service nodes in a consortium computing network. This method is applied to a distributed computing platform and includes the following steps: Upon receiving or responding to a computing task request submitted by a user, the system selects multiple candidate computing nodes from the associated node pool to execute the computing task and constructs a first node set. The associated node pool includes multiple computing nodes, which are associated with the distributed computing platform through pre-registration. Each computing node is a physical hardware device with computing capabilities. The service quality and response latency of each candidate computing node are obtained from historical interaction data, and node evaluation metrics are constructed based on service quality and response latency. Service quality is determined by the hardware configuration and model deployment version of the computing node and is used to measure the computing capability of the node. Response latency is used to measure the computing speed of the node. Node evaluation metrics are used to evaluate the quality of the computing node. Based on node evaluation metrics, an improved confidence upper bound algorithm is used to dynamically evaluate the quality of each candidate computing node. With the goal of maximizing the quality of computing nodes, a preset number of candidate computing nodes are selected from the first node set to form the second node set. The improved confidence upper bound algorithm is used to construct independent empirical mean and confidence upper bound index for service quality and response latency, respectively. Identify the relationships between candidate computing nodes in the second node set, and divide the second node set into multiple disjoint alliances based on the relationships; the relationships include one or more of the following: the geographical region where the node is located, the service provider to which it belongs, and the trust relationship; The system controls each candidate computing node in the second node set to collaboratively execute distributed computing tasks, and, under the premise of satisfying the node computing time budget constraint, uses a multi-agent reinforcement learning-based model to determine the computing contribution of each candidate computing node in the second node set in each time slot. Based on computational contribution, a two-level allocation mechanism is used to calculate the resource incentive value allocated to each candidate computing node. According to the resource incentive value, the scheduling priority of each candidate computing node in the distributed computing platform is dynamically updated to adjust the resource allocation strategy of subsequent computing tasks in the distributed computing platform. The first level of the two-level allocation mechanism allocates global resource incentives among all participating alliances based on the relative aggregate computational contribution of each alliance. The second level allocates the global incentives within each alliance based on the weighted individual computational contribution of member nodes.

[0006] Optionally, node evaluation metrics can be constructed based on service quality and response latency, including: Through calculation formula

[0007] Candidate computing nodes are obtained In the Wheel node evaluation indicators ;in, Indicates candidate computing nodes No. The normalized service quality score observed in each round; Indicates candidate computing nodes No. The response delay score observed in each round; This represents the weighting coefficient.

[0008] Optionally, the improved confidence upper bound algorithm constructs independent empirical means and confidence upper bound indices for service quality and response latency, respectively, including: Through calculation formula

[0009]

[0010] Candidate computing nodes are obtained No. upper confidence level of wheel service quality index ;in, Indicates as of the date Round of candidate computing nodes The average service quality score based on experience. This represents a confidence term used to measure the uncertainty of node estimation; This indicates the preset number of candidate computing nodes to be selected in each round; Indicates as of the date Round of candidate computing nodes The cumulative number of times selected; Through calculation formula

[0011] Candidate computing nodes are obtained No. upper confidence bound of wheel response delay ;in, Indicates as of the date The empirical average response latency score of candidate computing nodes in each round.

[0012] Optionally, a preset number of candidate computing nodes are selected from the first node set to form a second node set, including: Evaluation metrics for all candidate compute nodes in the first node set; Sort all candidate computing nodes in the first node set in descending order based on the values ​​of the node evaluation metrics. The top K candidate computing nodes are selected to form the second node set.

[0013] Optionally, the total resource incentive value obtained by each candidate computing node in the second node set collaboratively executing the distributed computing task. Incentive value based on basic resources With additional resource incentive value constitute; ;in, Basic resource incentive value Characterizes the inherent computational complexity or basic computing power requirement threshold of a distributed computing task. Additional resource incentive value Characterizes the marginal performance gain of the system resulting from the collaborative contribution of total computation time by multiple heterogeneous computing nodes; additional resource incentive value. The calculation expression is: ; This represents the total number of candidate computation nodes in the second node set; Indicates candidate computing nodes In the The computational contribution of each time slot characterizes the candidate computing node. In the The computation time contributed to the computation task in each time slot.

[0014] Optionally, the first level allocates global resource incentives among all participating alliances based on the relative aggregate contribution of each alliance. The second level allocates the global incentives within each alliance based on the weighted individual contribution of member nodes. The corresponding calculation formula is as follows:

[0015]

[0016]

[0017] in, Indicates alliance The aggregate computing contribution characterizes the computing consortium. The sum of effective total computing power and throughput output to the distributed computing system within a given time window; Indicates candidate computing nodes In the The computational contribution in each time slot; Indicates candidate computing nodes The empirical mean of the node evaluation index represents the candidate computing node. Statistical benchmarks for hardware reliability and network communication stability demonstrated during long-term operation; This represents the total number of alliances participating in the mission competition; Indicates the traversal number of the alliance; Indicates allocation to the alliance The global resource incentive value represents the macro-level system resource allocation quota issued by the distributed computing platform to various alliances; Indicates alliance The aggregated calculation contribution; Indicates candidate computation nodes after the second-level allocation. The final individual resource incentive value represents the candidate computing node. Micro-dynamic scheduling weights obtained within the cluster; Indicates alliance Candidate computing nodes In the The computational contribution in each time slot; Indicates candidate computing nodes The empirical mean of the node evaluation indicators.

[0018] Optionally, satisfying the node computation time budget constraint includes: The cumulative resource consumption of each candidate computing node across all time slots must not exceed the node's total available resource limit, and must satisfy the minimum aggregate computing threshold constraint; the specific mathematical constraints are as follows:

[0019]

[0020] in, Indicates the total time span. Indicates candidate computing nodes Provides the unit resource consumption required per unit of computation time; Indicates candidate computing nodes The total available resource limit is restricted throughout the entire time period; When all alliances are at The total computation time contributed within each time slot is lower than the set minimum aggregation computation threshold. When the computation task fails, the total resource incentive value for the current time slot is determined to be [value missing]. It is zero.

[0021] Optionally, a multi-agent reinforcement learning-based model is used to determine the computational contribution of each candidate computation node in the second node set in each time slot, including: Each candidate computing node is modeled as an autonomous intelligent agent, and a global state space, a joint action space, and an environmental feedback reward function are constructed. For the candidate computing nodes in the h-th time slot The local observation state vector of the node is obtained, which includes the computation time contribution of the previous time slot, the current basic resource incentive value, the minimum aggregation calculation threshold, the remaining available resource limit of the node, and the unit resource consumption. The local observation state vector is input into the multi-agent reinforcement learning model of the corresponding candidate computing node, and the target action is output through the model policy decision. The action space of the candidate computing node is a discrete action space, which represents the number of computing units that the node can submit in each time slot. The number of computing units is subject to the truncation limit of the ratio of the node's remaining available resource limit to the unit resource consumption. The computing contribution represents the number of computing units actually submitted by the node in each time slot. The environmental feedback reward function is calculated and determined by the individual resource incentive value finally obtained by the candidate computing nodes based on the two-level allocation mechanism, and is used to guide the multi-agent reinforcement learning model to update its decision-making strategy.

[0022] Optionally, the multi-agent reinforcement learning-based model specifically employs a multi-agent proximal policy optimization algorithm to solve for the computational contribution actions of candidate computation nodes; In the multi-agent proximal policy optimization algorithm, each candidate computation node contains an actor network optimized based on the first parameter and a critic network optimized based on the second parameter. The actor network maps the local observation state vector of the corresponding node to the action that calculates contribution, while the critic network is used to estimate the expected cumulative reward value of the corresponding local observation state.

[0023] Optionally, the actor network updates policy parameters by maximizing the objective function; the formula for calculating the objective function is as follows:

[0024]

[0025]

[0026]

[0027] in, The objective function of the actor network is represented, and the candidate computation nodes are characterized. Overall resource conversion efficiency during the execution of distributed tasks; Indicates candidate computing nodes Actor network parameters; This represents the expectation regarding multiple time slots. , Indicates the total number of time slots; This indicates taking the minimum value; This represents the probability ratio, reflecting the current computing power allocation strategy of a node when facing the current task. Compared with the old computing power allocation strategy The span between; Represents the clipping function; Indicates the clipping hyperparameter; if The value is ,if The value is ,if The value is ; This represents the estimated advantage function. A value greater than 0 indicates that the current resource scheduling action is more efficient than the previous average allocation scheme; Represents the local observation state and characterizes the candidate computation nodes. At any moment A real-time snapshot of the physical environment, including one or more of the node's remaining battery / budget, CPU / GPU idle rate, and minimum computing power threshold for the current task; Indicates candidate computing nodes At any moment Resource scheduling actions This represents the discount factor, reflecting the hardware depreciation rate of computing nodes or the urgency of monetizing resources. Higher-level representation nodes place greater emphasis on the sustained contribution of computing power over long-term future cycles. The lower the value, the more urgently the node needs to release idle computing power through the current task; Indicates the weighting parameter; Indicates the first The decay weight of the expected deviation of resource scheduling for each future time slot; Indicates candidate computing nodes From time Backwards The expected deviation in resource scheduling caused by each future time slot; Indicates candidate computing nodes At any moment The immediate reward obtained represents the candidate computing node. The individual resource incentive value ultimately obtained after the current computing power is invested and distributed through a two-level allocation mechanism; The state-value function represents the critic network's opinion on candidate computation nodes. Estimation of the value of the current state; This represents the critics' network's estimate of the long-term value of the next state; Indicates candidate computing nodes Comments with network parameters.

[0028] The present invention has at least the following beneficial effects: A first node set is constructed by receiving computing task requests and selecting candidate nodes with physical hardware computing capabilities from an associated node pool. Service quality determined by hardware configuration and response latency (a measure of computing speed) are used to construct node evaluation metrics based on historical interactions. An improved confidence upper bound algorithm, which calculates independent empirical means and confidence upper bound indices separately, is used to dynamically evaluate node quality, effectively overcoming the uncertainties caused by network and hardware performance fluctuations, and selecting a second node set with optimal composite performance. Furthermore, candidate computing nodes are divided into disjoint alliances based on attributes such as geographical region or service provider, while satisfying the total available resources of the nodes. Under the premise of equal physical constraints, a multi-agent reinforcement learning model is used to determine the computational contribution within the time slot. Finally, based on the computational contribution, a two-level allocation mechanism is used to allocate global resource incentives among the alliances according to the relative aggregate computational contribution, and within the alliance, the global incentives are further allocated according to the weighted individual computational contribution of the member nodes. This mechanism not only deeply couples the resource scheduling rules with the underlying hardware quota, avoiding the interference of subjective evaluation, but also effectively takes into account the macro-allocation efficiency between multi-physical alliances and the micro-scheduling accuracy of hardware nodes within the alliance, thereby significantly improving the throughput and stability of the entire heterogeneous distributed computing system. Attached Figure Description

[0029] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.

[0030] Figure 1 This is a workflow diagram of a distributed computing platform; Figure 2 This is a flowchart of a method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to an embodiment of the present invention; Figure 3 This is one of the implementation effect diagrams comparing the heterogeneous service node quality perception and reward allocation method for consortium computing power networks with the baseline method in another embodiment of the present invention; Figure 4 This is the second illustration comparing the implementation effect of the heterogeneous service node quality perception and reward allocation method for consortium computing power networks with the baseline method in another embodiment of the present invention. Figure 5 This is the third illustration comparing the implementation effect of the heterogeneous service node quality perception and reward allocation method for consortium computing power networks with the baseline method in another embodiment of the present invention; Figure 6 This is the fourth illustration comparing the implementation effect of the heterogeneous service node quality perception and reward allocation method for consortium computing power networks with the baseline method in another embodiment of the present invention; Figure 7This is the fifth illustration showing the implementation effect of the method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing power network compared with the baseline method in another embodiment of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0032] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0033] like Figure 1 As shown, in a typical distributed computing platform, the system usually processes large-scale computing tasks in parallel through a multi-stage physical pipeline. Specifically, its systematic workflow includes the following key operational mechanisms: First, after receiving a computing task request submitted by the front end (such as...), the platform... Figure 1 In stage ① of the process, the scheduler will perform dependency graph splitting and computational decoupling on the large task, forming multiple subtasks that can be executed in parallel (such as...). Figure 1 (Stage ② in the middle); subsequently, the system performs node quality evaluation and filtering selection from the underlying candidate physical computing node pool (such as... Figure 1 (Stage ③) Next, the selected target computing node cluster collaboratively processes the assigned subtasks (such as...) through the underlying network. Figure 1 (stage ④) and return their respective physical calculation results to the platform for data aggregation (e.g. Figure 1 (Stage ⑤) Finally, the system scheduling module will dynamically calculate and distribute resource incentive values ​​among the participating nodes based on the underlying computing power parameters (such as computation time and throughput) actually contributed by each node in this task. This is the quantitative allocation of subsequent scheduling priorities in the system (such as... Figure 1 (Stage 6) This workflow highlights two core technical decision points that fundamentally determine the throughput performance of a distributed system: first, a node selection strategy that identifies high-quality hardware resources; and second, a resource allocation and scheduling mechanism that ensures cross-node collaboration and is aware of the underlying computing power contribution.

[0034] However, when faced with the inherent heterogeneity of physical computing nodes in real-world network architectures, existing static or flattened scheduling schemes reveal significant shortcomings, severely limiting the overall performance of distributed systems. These shortcomings manifest in the following three aspects: Firstly, existing node selection mechanisms typically fail to jointly optimize service quality (such as computing power accuracy) caused by differences in hardware configuration and response latency caused by fluctuations in network communication conditions. Simple scheduling algorithms often overlook one aspect and cannot accurately reflect the comprehensive processing capabilities of computing nodes in real network environments.

[0035] Secondly, the underlying physical computing nodes are often limited by their own battery capacity, maximum allowable load, or heat dissipation limits, resulting in strict constraints on available time. Existing solutions often fail to couple these micro-hardware-level constraints with resource scheduling across multiple time slots, which can easily lead to local node overload and crashes or computing power allocation failures.

[0036] Thirdly, in real-world wide area networks (WANs), nodes are not completely isolated islands. They typically self-organize into network consortium topologies based on data centers of the same cloud service provider, geographical proximity, or trust protocols. Existing single-layer average resource allocation algorithms ignore this inherent hierarchical structure of system-consortium-node, failing to achieve macro-level global computing power balancing across multiple network consortia while simultaneously considering the performance differences among individual hardware nodes within a consortium. This coarse-grained scheduling leads to high-performance computing nodes failing to obtain matching resource scheduling weights, thereby reducing their willingness to maintain communication link stability and triggering a long-tail effect in processing latency across the entire heterogeneous cluster.

[0037] To address the network performance bottlenecks in existing distributed computing systems caused by unreasonable scheduling strategies when dealing with heterogeneous nodes, this invention proposes a quality awareness and reward allocation method for heterogeneous service nodes in consortium computing networks. This scheme aims to construct a quality-aware, two-level resource allocation mechanism adapted to the underlying network consortium topology. It adaptively evaluates the hardware service quality and network latency of nodes, and, under strict constraints on device physical resource limits, introduces a multi-agent reinforcement learning model to dynamically solve for the optimal computational contribution, thereby achieving precise resource scheduling and global energy efficiency optimization for heterogeneous computing systems in complex environments. The following specific embodiments illustrate the quality awareness and reward allocation method for heterogeneous service nodes in consortium computing networks provided by this invention.

[0038] Example 1 In this embodiment of the invention, the provided method for quality awareness and reward allocation of heterogeneous service nodes in a consortium computing network is applied to a distributed computing platform. For example... Figure 2 As shown, the method includes steps 21 to 26.

[0039] Step 21: Receive or respond to the computing task request submitted by the user, select multiple candidate computing nodes from the associated node pool to perform the computing task, and construct the first node set.

[0040] The associated node pool includes multiple computing nodes, which are associated with the distributed computing platform through pre-registration. Each computing node is a physical hardware device with computing capabilities, such as a rack server in an enterprise-level data center, an edge computing gateway deployed on the base station side, or an idle personal high-performance graphics workstation.

[0041] In practice, before connecting to the platform, each physical hardware device needs to install a dedicated system monitoring probe program at the operating system level. This probe program is responsible for collecting the device's underlying physical characteristics (such as CPU architecture and core count, GPU model and memory capacity, physical memory size, network card MAC address, and maximum physical bandwidth) and reporting these static hardware fingerprints to the platform's registration center. After successful platform verification, a globally unique node identifier is assigned to the hardware device, and the node is required to periodically send heartbeat packets to maintain its idle / online status, thus formally incorporating it into the associated node pool for unified physical resource management.

[0042] In one feasible implementation, when the platform frontend receives or responds to a user-submitted computing task request (e.g., a large language model distributed inference task) through the API gateway, the request message, in addition to containing a data payload, explicitly or implicitly carries basic hardware access tags (e.g., the task requires CUDA version 11.0 or higher and node GPU memory of at least 16GB). Upon receiving the request, the platform scheduler does not directly broadcast the task to the entire network, but first queries the real-time physical status table of the associated node pool. For example, the scheduler uses the basic hardware access tags in the request message to perform a first round of hard constraint filtering on all registered nodes in the pool, directly removing physical devices that are in a network disconnected state, have severely incompatible hardware specifications, or whose current CPU / GPU load has reached 100%. After this round of coarse-grained matching based on underlying hardware physical characteristics, the platform initially selects multiple candidate computing nodes from the associated node pool for executing the computing task (e.g., initially screening 50 physical servers that meet the basic operating conditions from 10,000 devices), thereby constructing the first node set. Suppose that the first node set is a set consisting of M candidate computing nodes whose performance characteristics need to be determined.

[0043] Step 22: Obtain the service quality and response latency of each candidate computing node from historical interaction data, and construct node evaluation indicators based on service quality and response latency.

[0044] After identifying a set of first nodes containing multiple basic physical execution capabilities, the distributed computing platform needs to further develop a deep, multi-dimensional capability profile of these physical devices in order to accurately assess their actual deliverable computing quality.

[0045] In practice, a distributed time-series database is deployed within the platform to continuously record the hardware consumption and network interaction logs of each physical device in the first node set when performing system tasks in the past. When the scheduler triggers an evaluation, it retrieves physical performance measurement data from this database for a preset time window (e.g., the past 24 hours or the most recent 100 task interactions).

[0046] In this embodiment of the invention, the quality of service is determined by the hardware configuration and model deployment version of the computing node, and is used to measure the node's computing power. The hardware configuration includes one or more of the following: the platform extracts the node's floating-point operation capability, video memory bandwidth utilization, and memory throughput from historical tasks. Different precision model deployment versions will result in hard differences in the fidelity or resolution of the output results.

[0047] In one feasible implementation, through calculation formula

[0048] Candidate computing nodes are obtained In the Normalized service quality score observed in round ;in, The theoretical accuracy coefficients of the model deployment version represent the candidate computing nodes. The specific computational model version loaded represents the upper limit of bit width and fidelity in mathematical terms. For example, for the same neural network, if a node is loaded with a full-precision floating-point version, this coefficient is 1.0; if a half-precision floating-point or integer quantized version is loaded due to resource constraints, this coefficient is objectively taken as 0.98 or 0.85 based on the proportion of feature loss caused by quantization pruning. This represents the execution overhead rate caused by hardware configuration, characterizing the candidate computing nodes. The actual physical computation loss that occurs when the hardware configuration supports this model version. It comprehensively reflects the physical limitations of the hardware, such as tensor slicing errors caused by insufficient video memory capacity or packet loss rate due to frequency reduction caused by thermal design power overload. The closer this value is to 0, the better the hardware configuration supports the current task. It represents the theoretical highest precision benchmark, which is a normalized constant. It characterizes the highest physical computing capability score limit allowed by the current distributed computing system platform under ideal conditions, and is used to map the final service quality score to the interval [0,1].

[0049] In this embodiment of the invention, response latency is used to measure the computing speed of nodes. Specifically, since nodes are distributed in different geographical locations or belong to different cloud service providers, network heterogeneity (such as routing hop count, bandwidth congestion, and physical distance) introduces unpredictable packet delays and long-tail effects. Therefore, response latency quantifies an end-to-end physical time: that is, the actual total time (usually in milliseconds ms or seconds) experienced from the platform sending subtask fragments to the target node through the network, to the node completing the processing, and then returning the full computation result to the platform master node through the network.

[0050] In one feasible implementation, through calculation formula

[0051]

[0052] Candidate computing nodes are obtained In the Response delay score observed in round ;in, Indicates relative latency; Indicates candidate computing nodes In the The response delay observed in the round; This indicates the preset minimum theoretical delay; This indicates the preset maximum theoretical latency. Response latency score. The value is [0, 1]. The closer the value is to 1, the shorter the round-trip time of the underlying network of the node, the faster the data delivery speed, and the better the communication link.

[0053] Step 23: Based on the node evaluation index, the quality of each candidate computing node is dynamically evaluated using an improved confidence upper bound algorithm. With the goal of maximizing the quality of computing nodes, a preset number of candidate computing nodes are selected from the first node set to form the second node set.

[0054] In real-world distributed heterogeneous computing networks, the operating status of physical devices is highly dynamic. For example, a high-performance GPU server may experience frequency reduction due to heat dissipation limitations caused by sustained high load, leading to a deterioration in its hardware service quality; simultaneously, network links may experience a surge in packet routing response latency due to sudden network congestion. Therefore, relying solely on a static, single evaluation cannot guarantee the long-term throughput efficiency of the cluster.

[0055] To address this issue, this invention improves upon the traditional confidence upper bound algorithm. The improved confidence upper bound algorithm constructs independent empirical means and confidence upper bound indices for service quality and response latency, respectively, and seeks a balance between utilizing known high-quality nodes and exploring insufficiently explored nodes.

[0056] Specifically, for each candidate computing node in the first node set, the platform scheduler maintains two independent historical empirical mean variables in the underlying state register: the empirical mean of the quality of service score. and the empirical mean of response latency score Among them, the empirical mean of service quality scores represents the results up to the [number]th [year]. Round of tasks, nodes The average hardware data processing fidelity or computational accuracy score demonstrated in various computational tasks. The empirical mean of the response latency score represents the performance up to the [number missing]. Round of tasks, nodes The average end-to-end network communication speed score demonstrated in each data packet transmission and reception.

[0057] To address the potential issue of insufficient sampling in physical probes, this invention calculates a dynamic confidence term. The calculation formula is as follows:

[0058] in, This indicates the preset number of candidate computing nodes to be selected in each round (e.g., K=20). Indicates the current global scheduling round; Indicates up to the number Round, candidate computing node The cumulative number of times a node is actually selected and issued physical operation commands. This formula measures the degree of a lack of knowledge about the underlying state of a particular node within the system. With each global round... The increase in edge devices means that if a device is not selected for task assignment for an extended period of time (i.e., ... Stagnation), confidence terms The numerical value will be mathematically amplified. This is equivalent to the system actively generating a probe stimulus signal, forcing the scheduler to allocate a small number of test tasks to the idle device in subsequent rounds to re-verify its current hardware and network health status.

[0059] Subsequently, through the calculation formula

[0060]

[0061] Candidate computing nodes are obtained No. upper confidence level of wheel service quality index and upper bound of confidence for response delay .

[0062] Finally, through the calculation formula

[0063] Candidate computing nodes are obtained In the Wheel node evaluation indicators ; This represents the weighting coefficient.

[0064] The platform's underlying memory queue is based on The values ​​are used to sort all nodes in descending order, and the IP and MAC addresses of the top K ranked physical nodes are directly extracted. This set of selected high-quality nodes constitutes the second set of nodes used to actually execute this concurrent computing task.

[0065] It should be noted that the advantage of this approach lies in the fact that by orthogonally decoupling service quality and response latency and independently constructing a confidence upper bound index, the system can clearly decompose the specific physical causes of node performance fluctuations in a mathematical model. When the network recovers, the previously recorded average value of high-quality hardware service is used as the basis for this approach. The node evaluation index calculation formula will immediately play a leading role again, enabling the high-performance node to quickly return to the first tier of scheduling. This technical solution of decoupling the detection of multi-dimensional underlying physical parameters and implementing dynamic evaluation is not obvious to those skilled in the art, and it greatly improves the stability and extreme value upper limit of the overall computing power output of distributed clusters in complex and fluctuating network environments.

[0066] Step 24: Identify the association relationships among the candidate computing nodes in the second node set, and divide the second node set into multiple disjoint alliances based on the association relationships.

[0067] In real-world distributed computing platforms, selected computing nodes are typically not completely independent, but rather physically or logically coupled. To effectively reduce overall system management overhead, share computing time budget constraints, and provide users with more stable service characteristics (such as lower intra-group communication latency), nodes often participate in distributed tasks in the form of alliances (or groups).

[0068] Specifically, after the distributed computing platform selects a second node set consisting of K candidate computing nodes from the candidate node pool, it identifies the relationships between the nodes by parsing the node's metadata and historical interaction logs. In this embodiment of the invention, the relationships include one or more of the following: the geographical region where the node is located, the service provider to which it belongs, and trust relationships.

[0069] Specifically, distributed computing platforms assess the geographical proximity of nodes by resolving their public IP addresses, routing hop counts, or registered physical locations. Nodes that are physically close to each other are typically deployed in the same edge computing center or adjacent network nodes, exhibiting extremely low network latency.

[0070] The service provider obtains this information by reading the identity authentication information when the node is registered. Specifically, this identifies which nodes are operated by the same computing power provider, deployed in the same data center, or located in the same local management domain.

[0071] Trust relationships are assessed based on pre-established trust and management protocols between nodes. For example, a group of nodes that have frequently collaborated on historical tasks and have not engaged in malicious behavior will be identified as having a high level of trust.

[0072] Based on the multi-dimensional relationships identified above, the platform employs corresponding clustering or label matching rules to guide the candidate computing nodes in the second node set to self-organize into m disjoint alliances. This alliance structure is formally defined as... Among these, all selected candidate computing nodes must belong to a certain alliance, i.e. , The union operation represents the set operation. This means that the total number of nodes in the alliance, when combined, is exactly equal to the second set of nodes. For example, suppose the second set of nodes... for Divided into , , ,but Furthermore, no computing node can belong to two or more alliances simultaneously; that is, for all... It will necessarily satisfy The intersection operation represents the set operation. To represent the empty set, , This indicates two different alliances.

[0073] After constructing a realistic alliance structure through the above steps, it will directly lay the foundation for subsequent complex two-level hierarchical reward distribution, thereby effectively solving the problem of non-stationarity in multi-agent interaction in heterogeneous distributed computing systems.

[0074] Step 25: Control each candidate computing node in the second node set to collaboratively execute the distributed computing task, and under the premise of meeting the node computing time budget constraint, use a multi-agent reinforcement learning-based model to determine the computing contribution of each candidate computing node in the second node set in each time slot.

[0075] In this embodiment of the invention, after the selection and alliance division of the second node set are completed, the task scheduling phase begins. When a user submits a distributed computing task, the platform decouples it into multiple subtasks that can be processed in parallel and distributes them to each candidate computing node for collaborative execution. The working time of the distributed system is divided into consecutive time slots. .

[0076] During this process, the system issues a total resource incentive value to each node for collaborative task execution. It is not a static constant, but rather determined by the basic resource incentive value. With additional resource incentive value constitute. .

[0077] Among them, the basic resource incentive value This represents the inherent computational complexity or basic computing power requirement threshold of a distributed computing task. It is preset by the user or platform.

[0078] Additional resource incentive value This characterizes the marginal performance gain of the system resulting from the collaborative contribution of total computation time by multiple heterogeneous computing nodes. The more computing power contributed collaboratively by the nodes, the higher the overall computational fidelity and response efficiency of the system. Additional resource incentive value. The calculation expression is: .in, This represents the total number of candidate computation nodes in the second node set; Indicates candidate computing nodes In the The computational contribution of each time slot characterizes the candidate computing node. In the The computation time contributed to the computation task in each time slot.

[0079] In determining the computational contribution of each time slot Previously, dual constraint verification was also implemented in embodiments of the present invention. Specifically, each node In total time span Within this period, the cumulative resource consumption must not exceed its physical limits or a predetermined budget. Mathematically, this can be expressed as: and .in, For nodes The unit resource consumption. This sets a limit on the total available resources. Simultaneously, to prevent task failure due to extreme computing power shortages, the total computing time contributed by all alliance members must reach a set threshold. .like If the task in the current time slot fails, the total resource incentive value for the current time slot will be reduced. Forced reset to zero.

[0080] During the execution phase of the distributed computing task, this invention models each candidate computing node in the second node set as an autonomous agent, and guides the agent to make resource allocation decisions by constructing a global state space, a joint action space, and an environmental feedback reward function.

[0081] To mitigate the inherent non-stationarity of multi-node interactions, this embodiment employs a multi-agent proximal policy optimization algorithm within the multi-agent Actor-Critic framework. In this architecture, each candidate computing node acts as an autonomous agent, comprising the following two core neural networks: Actor Network: Based on the first parameter Optimization is performed. Its core function is to optimize candidate computing nodes. The current local observation state vector is mapped to a specific action for calculating contribution.

[0082] During system training and dynamic scheduling, the actor network updates its policy parameters by maximizing the PPO objective function. For candidate computation nodes... The objective function of its actor network The calculation formula is as follows:

[0083]

[0084]

[0085]

[0086] in, The objective function of the actor network is represented, and the candidate computation nodes are characterized. Overall resource conversion efficiency during the execution of distributed tasks; Indicates candidate computing nodes Actor network parameters; This represents the expectation regarding multiple time slots. , Indicates the total number of time slots; This indicates taking the minimum value; This represents the probability ratio, reflecting the current computing power allocation strategy of a node when facing the current task. Compared with the old computing power allocation strategy The span between; Represents the clipping function; Indicates the clipping hyperparameter; if The value is ,if The value is ,if The value is ; This represents the estimated advantage function. A value greater than 0 indicates that the current resource scheduling action is more efficient than the previous average allocation scheme; Represents the local observation state and characterizes the candidate computation nodes. At any moment A real-time snapshot of the physical environment, including one or more of the node's remaining battery / budget, CPU / GPU idle rate, and minimum computing power threshold for the current task; Indicates candidate computing nodes At any moment Resource scheduling actions This represents the discount factor, reflecting the hardware depreciation rate of computing nodes or the urgency of monetizing resources. Higher-level representation nodes place greater emphasis on the sustained contribution of computing power over long-term future cycles. The lower the value, the more urgently the node needs to release idle computing power through the current task; Indicates the weighting parameter; Indicates the first The decay weight of the expected deviation of resource scheduling for each future time slot; Indicates candidate computing nodes From time Backwards The expected deviation in resource scheduling caused by each future time slot; Indicates candidate computing nodes At any moment The immediate reward obtained represents the candidate computing node The individual resource incentive value ultimately obtained after the current computing power is invested and goes through a two-level allocation mechanism; The state-value function represents the critic network's opinion on candidate computation nodes. Estimation of the value of the current state; This represents the critics' network's estimate of the long-term value of the next state; Indicates candidate computing nodes Comments with network parameters.

[0087] Candidate computing nodes Actor network parameters Update via gradient ascent:

[0088] in, This represents the learning rate of the actor network, used to control the step size for updating policy parameters. Indicates the network parameters of the actor Calculate the gradient.

[0089] Critic Network: Based on the second parameter Optimization is performed. Its core function is to estimate the expected cumulative reward value starting from the corresponding local observation state, which is used to evaluate the quality of the current state, thereby helping to guide the update direction of the actor network.

[0090] In this embodiment of the invention, the critic network is trained by minimizing the mean squared error between the predicted value and the target value. For candidate computation nodes... The loss function of its critic network Defined as:

[0091] in, Expressing expectations; Let represent the value function estimated by the critic network, and let represent the state from the local observations. The expected cumulative return value upon departure; The target reward represents the benchmark used during network training.

[0092] Critics Network Parameters Update via gradient descent:

[0093] in, It is the learning rate of the critic network, used to control the update step size of the value function parameters. Indicates the network parameters of the critics Calculate the gradient.

[0094] Specifically, in the first A time slot, a node The sensor acquires the local observation state vector .in, This indicates the historical calculation contribution of each node in the previous time slot; This represents the base resource incentive value for the current distributed task. This is the minimum aggregation calculation threshold for this task; and These represent the current remaining available resource limit (remaining budget) and the unit resource consumption (operating cost) of the node, respectively.

[0095] Based on observation status ,node The action space is defined as the discrete action space. This indicates the current proposed computational contribution (number of computational units or computation time of the contribution). To strictly adhere to resource quota constraints, a truncation restriction is imposed on the action space during implementation: if the number of output computational units exceeds the limit, it is forcibly truncated to the ratio of the remaining available resource quota of that node to the unit resource consumption. Within the largest integer boundary of ). It should be noted that, in this embodiment of the invention, the number of computing units is the strategy choice made by the model in the discrete action space, while the computation time is the physical mapping and execution result of this strategy choice in the distributed computing system. Logically, the two are interchangeable and equivalently convertible, both representing the computational contribution.

[0096] The goal of each candidate computing node is to maximize its reward. This reward function is the individual resource incentive value ultimately obtained from the aforementioned two-tier allocation mechanism (first allocated to the consortium, then allocated to the nodes themselves). The reward is determined directly through calculation and serves as the reward signal fed back from the environment to the node. The reward function is expressed as: .

[0097] Through the aforementioned network interaction and dynamic updates, the heterogeneous distributed computing system can learn the optimal computing contribution strategy for each autonomous computing node under constantly changing budget and computing power requirements.

[0098] Step 26: Based on the computational contribution, calculate the resource incentive value allocated to each candidate computing node through a two-level allocation mechanism, and dynamically update the scheduling priority of each candidate computing node in the distributed computing platform according to the resource incentive value, so as to adjust the resource allocation strategy of subsequent computing tasks in the distributed computing platform.

[0099] After the computing nodes collaboratively complete the distributed task for the current time slot, the total resource incentive value generated needs to be settled and allocated. A simple average allocation scheme would ignore the heterogeneity of the nodes, leading to insufficient compensation for high-performance nodes. Therefore, this embodiment employs a two-level allocation mechanism that enhances quality differentiation while ensuring internal fairness, and uses the result as a feedback signal for subsequent scheduling decisions.

[0100] In this embodiment of the invention, the first level of the two-level allocation mechanism allocates global resource incentives among all participating alliances based on the relative aggregated contribution of each alliance.

[0101] Specifically, in the first level, the alliance is treated as a unified entity, and global resource incentives are allocated among all participating alliances based on the relative aggregate contribution of each alliance.

[0102] First, the computing consortium Aggregate calculation contribution To differentiate between high-quality and low-quality alliances, this embodiment of the invention employs a secondary metric to quantify alliance contributions:

[0103] in, Indicates alliance The aggregate computing contribution characterizes the computing consortium. The sum of effective total computing power and throughput output to the distributed computing system within a given time window; Indicates candidate computing nodes In the The computational contribution in each time slot; Indicates candidate computing nodes The empirical mean of the node evaluation index represents the candidate computing node. Statistical benchmarks for hardware reliability and network communication stability demonstrated during long-term operation.

[0104] Subsequently, the allocation to the alliance was calculated based on the relative contribution ratio. Global resource incentive value :

[0105] in, This represents the total number of alliances participating in the mission competition; Indicates the traversal number of the alliance; Indicates allocation to the alliance The global resource incentive value represents the macro-level system resource allocation quota issued by the distributed computing platform to various alliances; Indicates alliance The aggregated calculation contribution.

[0106] Based on the above calculations, it exhibits high stability ( ) and high computing power investment ( The alliance will receive an exponentially amplified allocation of system resources.

[0107] In the league Obtain its global resource quota Then, the second-level allocation begins. In this embodiment of the invention, the second level allocates the global incentives within each alliance based on the weighted individual contribution of member nodes.

[0108] Specifically, through the calculation formula

[0109] After obtaining the candidate computation nodes in the second-level allocation The final individual resource incentive value obtained Characterize candidate computing nodes Micro-dynamic scheduling weights obtained within the cluster Indicates alliance Candidate computing nodes In the The computational contribution in each time slot; Indicates candidate computing nodes The empirical mean of the node evaluation indicators.

[0110] In this embodiment of the invention, the resource incentive values ​​for each candidate computing node are completed. After calculation, the distributed computing platform treats this as an important feedback signal, used to dynamically update the scheduling priority of each node within the platform. In one feasible implementation, the specific logic is as follows: The platform maintains a global candidate node scheduling weight matrix. When a node... Calculated system incentive evaluation value When the value remains consistently high, it objectively indicates that the node not only has stable hardware computing power output (no downtime, high throughput), but also that its alliance's network communication link is excellent (low response latency, low packet loss rate). At this point, this quantified value ( The parameters are converted into control parameters and added to the historical scheduling weights of the node to complete the dynamic iteration of the weights.

[0111] When a distributed computing platform receives new constraints with higher computational complexity or stricter latency (i.e., corresponding to higher basic resource incentive values), When performing critical computational tasks, the updated scheduling weight matrix will be read. At this point, the invention no longer relies solely on the confidence upper bound algorithm for blind computational power exploration. Instead, it prioritizes routing task data packets and computation instructions to nodes or alliances with high scheduling priority from the underlying network topology, thereby significantly reducing the system's end-to-end task processing latency. Conversely, for alliances that consistently produce extremely low effective computational power during collaborative execution, frequently causing their alliances to fail to reach the minimum aggregate computation threshold, the updated scheduling weight matrix will be read. or evaluation indicators Nodes consistently falling below a preset safety threshold will trigger a hardware degradation instruction, dynamically lowering their scheduling priority. In subsequent concurrent task distribution streams, these nodes will be isolated from the core computing chain and will instead only receive latency-insensitive background asynchronous tasks. Through this closed-loop control mechanism based on quantitative feedback, this invention achieves dynamic load balancing of the underlying distributed heterogeneous computing power pool and global maximization of hardware resource utilization.

[0112] Example 2 To verify the objective effectiveness, robustness, and technical advantages of the proposed heterogeneous service node quality perception and reward allocation method for consortium computing power networks (hereinafter referred to as the QARC framework) under multi-index optimization, this embodiment built a simulation experimental environment and conducted a comprehensive comparative experimental evaluation using real-world datasets.

[0113] The simulation evaluation in this embodiment is divided into two main stages: the node selection mechanism evaluation stage and the reward distribution mechanism evaluation stage.

[0114] In the node selection mechanism evaluation phase, to realistically simulate objective quality of service (QoS) changes and network topology heterogeneity in a distributed computing system, this embodiment uses two real open-source datasets: the AMF dataset and the NetLatency dataset. The AMF dataset contains real QoS records of 142 terminals calling 4500 Web services over 64 time slices. This embodiment uses the throughput from this dataset as the underlying physical metric for measuring node service quality, calculates the average performance of each node, and normalizes it to the interval [0,1], which serves as the normalized service quality score in Example 1. The NetLatency data selected is the Seattle real network trajectory data, which records the round-trip latency of 99 network nodes over 688 time slices. This embodiment normalizes the average RTT to the interval [0,1] and defines the response latency score as... , This represents the normalized time delay.

[0115] In the basic parameter settings, the total number of candidate computing nodes M varies from {30, 40, 50, 60, 70} (default is M=50); the number of nodes to be selected in each round K varies from {10, 20, 30, 40, 50} (default is K=20); and the statistical mean of the balancing weight coefficient varies from {0.2, 0.4, 0.6, 0.8, 1.0} (default is 0.6). In each round, the real-time utility score of a node is dynamically and randomly generated based on the above true mean and combined with a Gaussian distribution.

[0116] In this embodiment, the top 20 nodes with the highest evaluation metrics during the node selection phase are selected as the final computing nodes participating in the game. The physical characteristics and benchmark values ​​of the task are derived from a real large-scale data center cluster load dataset—the Alibaba Cluster Trace dataset.

[0117] Assuming the basic reward of the task Positively correlated with the number of instances contained in each task, and normalized to the interval [0,1]. The average total budget limit B per node is dynamically selected from {10,20,30,40,50} (default is B=50); the operating cost per unit of computation time provided by a node. Dynamically selected from {10, 12, 14, 16, 18} (default value is c=10); the system-defined minimum aggregation calculation threshold for the task. Dynamically select from {0.6, 1.2, 1.8, 2.4, 3.0} (default setting is...) Θ =0.6). The number of self-organized computing alliances is preset to 5, and the optional action space of each candidate computing node in each time slot is discretized as {0,0.1,0.2,...,1}.

[0118] To comprehensively demonstrate the technological advancement of this invention, several representative industry baseline algorithms were introduced for comparison in both phases of evaluation. Specifically, the comparison algorithms in the node selection phase included the theoretically optimal strategy algorithm (Optimal), the pure greedy exploitation strategy (Greedy), and the fixed probability exploration strategy (…). -Greedy), decay probability exploration strategy ( -Decreasing) and purely random selection strategies. Comparative algorithms in the reward allocation phase include: Multi-Agent Q-Learning (MAQL), ​​short-sighted greedy resource strategy, and conservative risk-averse strategy.

[0119] By running for 10 5 The convergence, cumulative node evaluation index, and cumulative regret index of the proposed node selection algorithm (NSA) were quantitatively analyzed through rounds of iterative simulation. The results are as follows: Figure 3 , Figure 4 , Figure 5 as well as Figure 6 As shown.

[0120] like Figure 3 As shown, for all algorithms, the node evaluation metric increases with the number of rounds, reaching a maximum of 10. 5 In the final round, NSA consistently outperformed all four baselines and came very close to the Optimal strategy. In the final round, NSA achieved approximately 49.9% and 12.8% higher satisfaction rates compared to Random and Greedy, respectively, reaching 97.97% of the theoretical optimum. Figure 4 and Figure 5The impact of the number of selected nodes K on the number of candidate nodes M is demonstrated. Node evaluation metrics generally increase with increasing K. When M increases from 30 to 70, the performance gap between Greedy and NSA widens, while NSA remains robust, illustrating its effectiveness in large-scale heterogeneous environments. Figure 6 Further demonstration of preference weights The impact. In different At these values, NSA consistently outperforms baseline methods and remains close to the Optimal strategy.

[0121] like Figure 7 As shown, in long-term scheduling evaluation, when the average initial resource limit (budget B) is set to 50, the scheduling method proposed in this invention achieves a long-term average reward increase of 30.8% and 101.7% respectively compared to the Conservative baseline and the Random baseline. This fully demonstrates that the Actor-Critic dual network in the model of this invention has successfully learned how to strategically allocate computing resources across time slots under a limited budget through long-term evaluation of the advantage function, avoiding the drawbacks of excessive consumption in the early stage leading to no computing power available in the later stage, or excessive conservatism in the early stage leading to missing out on high immediate rewards.

[0122] It is evident that by receiving computing task requests and selecting candidate nodes with physical hardware computing capabilities from the associated node pool to construct the first node set, and by constructing node evaluation metrics based on service quality determined by hardware configuration and response latency measuring computing speed through historical interactions, and by dynamically evaluating node quality using an improved confidence upper bound algorithm that calculates independent empirical means and confidence upper bound indices respectively, the uncertainty caused by network and hardware performance fluctuations can be effectively overcome, thus selecting the second node set with optimal composite performance. Furthermore, by dividing candidate computing nodes into disjoint alliances based on attributes such as geographical region or service provider, while ensuring the total available resources of the nodes are met... Under physical constraints such as source quotas, a multi-agent reinforcement learning model is used to determine the computational contribution within a time slot. Finally, based on this computational contribution, a two-level allocation mechanism is used to allocate global resource incentives among alliances according to their relative aggregate computational contribution, and within the alliance, the global incentives are further allocated based on the weighted individual computational contribution of member nodes. This mechanism not only deeply couples resource scheduling rules with underlying hardware quotas, avoiding interference from subjective evaluations, but also effectively balances the macro-allocation efficiency among multi-physical alliances with the micro-scheduling accuracy of hardware nodes within the alliance, thereby significantly improving the throughput and stability of the entire heterogeneous distributed computing system.

[0123] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

[0124] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.

Claims

1. A method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network, characterized in that, Applied to a distributed computing platform, the method includes: Upon receiving or responding to a computing task request submitted by a user, the system selects multiple candidate computing nodes from an associated node pool to execute the computing task, thus constructing a first node set. The associated node pool includes multiple computing nodes, which are associated with the distributed computing platform through pre-registration, and each computing node is a physical hardware device with computing capabilities. The service quality and response latency of each candidate computing node are obtained from historical interaction data, and a node evaluation index is constructed based on the service quality and response latency. The service quality is determined by the hardware configuration and model deployment version of the computing node and is used to measure the node's computing power. The response latency is used to measure the node's computing speed. The node evaluation index is used to evaluate the quality of the computing node. Based on the node evaluation metrics, an improved confidence upper bound algorithm is used to dynamically evaluate the quality of each candidate computing node. With the goal of maximizing the quality of computing nodes, a preset number of candidate computing nodes are selected from the first node set to form a second node set. The improved confidence upper bound algorithm is used to construct independent empirical mean and confidence upper bound index for the service quality and the response latency, respectively. Identify the association relationships among candidate computing nodes in the second node set, and divide the second node set into multiple disjoint alliances based on the association relationships; the association relationships include one or more of the following: the geographical region where the node is located, the service provider to which it belongs, and the trust relationship; Controlling each candidate computing node in the second node set to collaboratively execute distributed computing tasks, and under the premise of satisfying the node computing time budget constraint, using a multi-agent reinforcement learning-based model to determine the computing contribution of each candidate computing node in the second node set in each time slot; the step of using a multi-agent reinforcement learning-based model to determine the computing contribution of each candidate computing node in the second node set in each time slot includes: Each candidate computing node is modeled as an autonomous intelligent agent, and a global state space, a joint action space, and an environmental feedback reward function are constructed. For the Candidate computing nodes in each time slot The local observation state vector of the node is obtained, which includes the computation time contribution of the previous time slot, the current basic resource incentive value, the minimum aggregation calculation threshold, the remaining available resource limit of the node, and the unit resource consumption. The local observation state vector is input into the multi-agent reinforcement learning model of the corresponding candidate computing node, and the target action is output through the model policy decision. The action space of the candidate computing node is a discrete action space, which represents the number of computing units that the node can submit in each time slot. The number of computing units is subject to the truncation limit of the ratio of the node's remaining available resource limit to the unit resource consumption. The computing contribution represents the number of computing units actually submitted by the node in each time slot. The environmental feedback reward function is calculated and determined by the individual resource incentive value finally obtained by the candidate computing node based on the two-level allocation mechanism, and is used to guide the multi-agent reinforcement learning model to update the decision strategy. Based on the computational contribution, a two-level allocation mechanism is used to calculate the resource incentive value allocated to each candidate computing node. According to the resource incentive value, the scheduling priority of each candidate computing node in the distributed computing platform is dynamically updated to adjust the resource allocation strategy of subsequent computing tasks in the distributed computing platform. The first level of the two-level allocation mechanism allocates global resource incentives among all participating alliances based on the relative aggregate computational contribution of each alliance. The second level allocates the global incentives within each alliance based on the weighted individual computational contribution of member nodes.

2. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 1, characterized in that, The node evaluation metrics constructed based on the service quality and the response latency include: Through calculation formula Candidate computing nodes are obtained In the Wheel node evaluation indicators ;in, Indicates candidate computing nodes No. The normalized service quality score observed in each round; Indicates candidate computing nodes No. The response delay score observed in each round; This represents the weighting coefficient.

3. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 2, characterized in that, The improved confidence upper bound algorithm constructs independent empirical means and confidence upper bound indices for the service quality and the response latency, respectively, including: Through calculation formula Candidate computing nodes are obtained No. upper confidence bound index of wheel service quality ;in, Indicates as of the date Round of candidate computing nodes The average service quality score based on experience. This represents a confidence term used to measure the uncertainty of node estimation; This indicates the preset number of candidate computing nodes to be selected in each round; Indicates as of the date Round of candidate computing nodes The cumulative number of times selected; Through calculation formula Candidate computing nodes are obtained No. upper confidence bound of wheel response delay ;in, Indicates as of the date The empirical average response latency score of candidate computing nodes in each round.

4. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 3, characterized in that, The step of selecting a preset number of candidate computing nodes from the first node set to form a second node set includes: The evaluation metrics for all candidate computing nodes in the first node set; Sort all candidate computing nodes in the first node set in descending order according to the values ​​of the node evaluation index. The top K candidate computing nodes are selected to form the second node set.

5. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 4, characterized in that, The total resource incentive value obtained by each candidate computing node in the second node set collaboratively executing the distributed computing task. Incentive value based on basic resources With additional resource incentive value constitute; ;in, Basic resource incentive value Characterizes the inherent computational complexity or basic computing power requirement threshold of a distributed computing task. Additional resource incentive value Characterizes the marginal performance gain of the system resulting from the collaborative contribution of total computation time by multiple heterogeneous computing nodes; additional resource incentive value. The calculation expression is: ; This represents the total number of candidate computation nodes in the second node set; Indicates candidate computing nodes In the The computational contribution of each time slot characterizes the candidate computing node. In the The computation time contributed to the computation task in each time slot.

6. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 5, characterized in that, The first level allocates global resource incentives among all participating alliances based on the relative aggregate contribution of each alliance. The second level allocates the global incentives within each alliance based on the weighted individual contribution of member nodes. The corresponding calculation formula is as follows: in, Indicates alliance The aggregate computing contribution characterizes the computing consortium. The sum of effective total computing power and throughput output to the distributed computing system within a given time window; Indicates candidate computing nodes In the The computational contribution in each time slot; Indicates candidate computing nodes The empirical mean of the node evaluation index represents the candidate computing node. Statistical benchmarks for hardware reliability and network communication stability demonstrated during long-term operation; This represents the total number of alliances participating in the mission competition; Indicates the traversal number of the alliance; Indicates allocation to the alliance The global resource incentive value represents the macro-level system resource allocation quota issued by the distributed computing platform to various alliances; Indicates alliance The aggregated calculation contribution; Indicates candidate computation nodes after the second-level allocation. The final individual resource incentive value represents the candidate computing node. Micro-dynamic scheduling weights obtained within the cluster; Indicates alliance Candidate computing nodes In the The computational contribution in each time slot; Indicates candidate computing nodes The empirical mean of the node evaluation indicators.

7. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 6, characterized in that, The constraint of satisfying the node computation time budget includes: The cumulative resource consumption of each candidate computing node across all time slots must not exceed the node's total available resource limit, and must satisfy the minimum aggregate computing threshold constraint; the specific mathematical constraints are as follows: in, Indicates the total time span. Indicates candidate computing nodes Provides the unit resource consumption required per unit of computation time; Indicates candidate computing nodes The total available resource limit is restricted throughout the entire time period; When all alliances are at The total computation time contributed within each time slot is lower than the set minimum aggregation computation threshold. If the computation task fails, the total resource incentive value for the current time slot is determined to be [value missing]. It is zero.

8. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 7, characterized in that, The multi-agent reinforcement learning-based model specifically employs a multi-agent proximal policy optimization algorithm to solve for the computational contribution actions of candidate computation nodes. In the multi-agent proximal policy optimization algorithm, each candidate computing node includes an actor network optimized based on the first parameter and a critic network optimized based on the second parameter. The actor network maps the local observation state vector of the corresponding node to the action of calculating contribution, and the critic network is used to estimate the expected cumulative reward value of the corresponding local observation state.

9. The method for quality perception and reward allocation of heterogeneous service nodes in a consortium computing network according to claim 8, characterized in that, The actor network updates its policy parameters by maximizing an objective function; the formula for calculating the objective function is as follows: in, The objective function of the actor network is represented, and the candidate computation nodes are characterized. Overall resource conversion efficiency during the execution of distributed tasks; Indicates candidate computing nodes Actor network parameters; This represents the expectation regarding multiple time slots. , Indicates the total number of time slots; This indicates taking the minimum value; This represents the probability ratio, reflecting the current computing power allocation strategy of a node when facing the current task. Compared with the old computing power allocation strategy The span between; Represents the clipping function; Indicates the clipping hyperparameter; if The value is ,if The value is ,if The value is ; This represents the estimated advantage function. A value greater than 0 indicates that the current resource scheduling action is more efficient than the previous average allocation scheme; Represents the local observation state and characterizes the candidate computation nodes. At any moment A real-time snapshot of the physical environment, including one or more of the node's remaining battery / budget, CPU / GPU idle rate, and minimum computing power threshold for the current task; Indicates candidate computing nodes At any moment Resource scheduling actions; This represents the discount factor, reflecting the hardware depreciation rate of computing nodes or the urgency of monetizing resources. Higher-level representation nodes place greater emphasis on the sustained contribution of computing power over long-term future cycles. The lower the value, the more urgently the node needs to release idle computing power through the current task; Indicates the weighting parameter; Indicates the first The decay weight of the expected deviation of resource scheduling for each future time slot; Indicates candidate computing nodes From time Backwards The expected deviation in resource scheduling caused by each future time slot; Indicates candidate computing nodes At any moment The immediate reward obtained represents the candidate computing node The individual resource incentive value ultimately obtained after the current computing power is invested and goes through a two-level allocation mechanism; The state-value function represents the critic network's opinion on candidate computation nodes. Estimation of the value of the current state; This represents the critics' network's estimate of the long-term value of the next state; Indicates candidate computing nodes The network parameters of the commentators.

Images